empes

Case study · 2026

Building a consistent AI character system for branded content

An e-commerce brand was making product visuals with AI, but every batch looked different. We built a repeatable workflow, so the output stays consistent and the team runs it on its own.

Client
E-commerce
Status
Shipped
Service
Empes AI
Scope
AI, workflow design, visual consistency

Problem

The client was producing product visuals with AI generation tools, but the quality of the output and the consistency of the character changed from campaign to campaign. Each session gave different results: different lighting, different proportions, a different style and brand feel. The team spent more time fixing and choosing outputs than producing content. There was no system, only repeated trial and error.

Underneath was a question of scale. The brand wanted video for far more of its catalogue than the hero products, to compete on marketplace product pages and in social placements, and traditional production could not cover that at a sensible cost and speed. AI could help, but only if the output was consistent and usable.

Approach

The problem was not the generation tool. It was the lack of a structured way of preparing what goes into it.

We started by auditing how generation worked: which prompts were used, how assets were prepared, and what the review criteria were. From there we designed a reusable workflow built on consistent inputs, structured prompts and a clear review step. The goal was to make consistency the default, not the exception.

From trial and error to a systemTwo columns. Before: each session started from scratch and gave different results, which the team fixed and chose by hand. After: prepared inputs and templates, generation, and a review against written criteria.BeforeA new prompt each sessionGenerateDifferent lighting, proportions and styleFix and choose by handEvery session different. More time selectingthan producing.AfterPrepared reference assets and prompttemplatesGenerateReview against written criteriaReady for marketplace and social formatsConsistency by default, run by the team itself.control theinputsFrom trial and error to a systemTwo columns. Before: each session started from scratch and gave different results, which the team fixed and chose by hand. After: prepared inputs and templates, generation, and a review against written criteria.BeforeA new prompt each sessionGenerateDifferent lighting, proportions and styleFix and choose by handEvery session different. More time selecting thanproducing.control the inputsAfterPrepared reference assets and prompt templatesGenerateReview against written criteriaReady for marketplace and social formatsConsistency by default, run by the team itself.
Read the diagram as text
  1. Before: A new prompt each session, then Generate, then Different lighting, proportions and style, then Fix and choose by hand. Every session different. More time selecting than producing.
  2. control the inputs
  3. After: Prepared reference assets and prompt templates, then Generate, then Review against written criteria, then Ready for marketplace and social formats. Consistency by default, run by the team itself.

Implementation

The generation workflow, with its review loopFive steps: prepared inputs, generation, review against written criteria, the asset library, and marketplace and social formats. Outputs that fail review are regenerated, with the reason recorded.Preparedinputsreference assetsandparameterisedpromptsGenerationComfyUIReviewagainst writtencriteriaFails:regenerate, andrecord whyAsset libraryfolder andnaming rulesReady topublishmarketplace andsocial formatsThe generation workflow, with its review loopFive steps: prepared inputs, generation, review against written criteria, the asset library, and marketplace and social formats. Outputs that fail review are regenerated, with the reason recorded.Prepared inputsreference assets andparameterised promptsGenerationComfyUIReviewagainst written criteriaFails:regenerate, andrecord whyAsset libraryfolder and naming rulesReady to publishmarketplace and social formats
Inputs are controlled before anything is generated. Documentation and training mean the team runs the whole loop without us.
Read the diagram as text
  1. Prepared inputs (reference assets and parameterised prompts)
  2. Generation (ComfyUI)
  3. Review (against written criteria). Otherwise: Fails: regenerate, and record why
  4. Asset library (folder and naming rules)
  5. Ready to publish (marketplace and social formats)
  • Prompt architecture. Prompts that are structured, reusable and parameterised.
  • Character consistency. Prepared reference assets and controls on the inputs.
  • Workflow documentation. The generation and review process, step by step.
  • Asset organisation. A folder structure and naming conventions.
  • Review criteria. What passes, what gets regenerated, and why.
  • Handoff and training. So the team can run the workflow on its own.

Outcome

  • Less time spent choosing outputs and correcting them by hand.
  • More consistent visuals from one campaign batch to the next.
  • The team moved from ad hoc generation to a structured, repeatable process.
  • The workflow is reused across product lines without starting from scratch each time.
  • Output is prepared in the formats marketplace and social placements need.

What this shows

Consistency in AI-generated content is rarely just a prompting problem. It is a systems problem, solved by structuring inputs, controls and review before the generation step begins. Once that system exists, the team can produce more without losing the look that makes it theirs.

If your AI output looks different every time, book a call and show us a batch.

Stack
ComfyUI, LLM, Python
Type
Strategy + implementation
  • ai
  • workflow-design
  • ai-characters
  • visual-consistency
  • branded-content
  • e-commerce

Book a 30-min scoping call

Bring one process. We will tell you whether automation changes it, and what it would take.

Book a 30-min scoping call hello@empes.it